E—A Language for Thread-Level Parallel Programming on Synchronous Shared Memory NOCs

As systems on chip are evolving to networks on chip (NOC) providing a unified communication infrastructure for a number of computational resources, being able to easily implement computational tasks as a parallel program that can be efficiently executed by multiple resources together is becoming increasingly important. Recent advances in thread-level parallel (TLP) architectures have made it possible to implement efficiently an easy-to-use synchronous shared memory programming model (Parallel Random Access Machine, PRAM) on a NOC. In this paper we describe a novel programming language, called e, for fine-grained TLP programming on synchronous shared memory NOC architectures realizing the PRAM model. The language uses a familiar c-like syntax and provides support for shared and private variables, arbitrary hierarchical groups of threads, and synchronous control structures. This allows a programmer to use various advanced TLP programming techniques like data parallelism, divide-and-conquer technique, different blocking techniques, and both synchronous and asynchronous programming style. We will also shortly experiment the e-language with real parallel programs using our experimental e-compiler and scalable Eclipse NOC architecture. Key-Words: Thread-level parallel programming languages, parallel random access machine, networks on chip ing between threads making MPAs difficult to program. In shared memory architectures (SMA) communication happens just by referring to the shared memory, data and program partitioning happens by allocating data structures from the shared memory making programming much easier. Unfortunately most SMAs consist of multiple interconnected processor-cache pairs making cache coherency and synchronicity maintenance very expensive. Recent advances in TLP architectures [3, 4], however, suggest that it possible to implement a cacheless and synchronous SMA realizing the parallel random access machine (PRAM) model [5] on a NOC. In this paper we describe a novel programming language, called e, for fine-grained TLP programming on synchronous SMA NOCs realizing the PRAM model. The language uses a familiar c-like syntax and provides support for shared and private variables, arbitrary hierarchical groups of threads, and synchronous control structures. This allows a programmer to use various advanced TLP programming techniques like data parallelism, divide-and-conquer techniques, different blocking techniques, and both synchronous and asynchronous programming style [6]. We made also a short experimentation with the e-language by implementing a number of familiar TLP algorithms with e, compiling them with our experimental e-compiler and executing them in our scalable high performance Eclipse NOC architecture [4].